Good morning,

This week we cover what Anthropic's prospectus tells the market, including a warning that its worst case may be too big for anyone to insure. Plus the quick bites, the interview and a company to keep in mind: Harbor, which offers insurance for AI agents.

But first, a follow-up on last edition's story about AI and cybersecurity.

Last edition, we discussed how AI was starting to have a lollapalooza effect on cybersecurity: by cutting the (brain and compute) cost of compromising a system, it lets hackers move faster and go after a much broader range of targets (like French supermarkets’ fridges).

Barely a few days after we published, an unusual run of incidents led security experts to call the month “Black October”. Some of the most problematic incidents:

So what is happening? To understand it, I believe the most important thing is the difference between a vulnerability and an exploit. A vulnerability is like a crack in a wall: you know it is there and that it is a problem, but it does not bring your house down, and hopefully you have time to patch it before it grows. An exploit is a way of turning that crack into a real problem, like filling it with TNT until the wall gives way. A house has many small cracks and fixing them all takes time and effort, so for years security experts assessed the risk of each vulnerability and patched the most dangerous first.

AI is reshaping this on both sides. First, that approach assumed the people patching vulnerabilities would be quicker than the people building exploits; AI has drastically sped up both, and who is faster is no longer clear. Second, it relied on experts judging which vulnerabilities mattered, but AI now finds “innovative” ways, unthought of by humans, to turn them into exploits, so yesterday’s “low-risk vulnerabilities” can become very dangerous ones.

This is exactly what happened with PHP two weeks ago. The exploit relied on two vulnerabilities discovered in 2004 and 2019, judged too small to be a priority at the time and therefore never fixed. In August, a security firm using OpenAI’s GPT-5.6 Sol found a way to combine them into an exploit that lets an attacker run almost any code they wish on a website’s server (i.e. a big issue). They reported the exploit to the PHP team, but before the fix was ready the exploit leaked online, giving anyone the ability to take control of a website running PHP (i.e. up to 70% of the web).

This kind of event should prompt all of us, companies and individuals alike, to understand what has changed and what we should do to protect ourselves. See an example at the end of this newsletter, with Harbor.

After months of waiting and rumours, Anthropic finally made its S-1 public on 26 October, the first frontier lab to do so. Aside from the unbelievable growth that would make anyone green with envy, the 400-plus-page document highlights some of the company's main fears and opportunities.

First, the money. Quarterly revenue went from $376m in early 2025 to $11.6bn in Q2 2026, and Anthropic told investors it reached $16.8bn in Q3. That single quarter is almost four times what the company made in all of 2025, and at that pace it earns its entire 2023 revenue in under six hours. Customers spending more than $100,000 a year went from about 950 to over 11,900 in twelve months, and with roughly 5,300 staff, Anthropic now makes $6.6m of annual revenue per employee. The growth is also turning into profit: compute costs 55 cents per dollar of revenue, against $2.01 a year earlier, and the second quarter brought a first operating profit of $974m. The $80.7bn first-half loss is mostly accounting, from revaluing convertible notes as the valuation soared. The real weight lies ahead: about $417bn of new compute commitments signed since January, payable “regardless of actual utilization”, plus another $125–150bn for capacity arriving in 2028–31.

Second, the risks, which fill about 94 pages. The opening summary warns that AI is already used for "high-profile cyberattacks" and that misuse in biology "could result in even greater harm". Anthropic admits its models "can function as a general-purpose tool for gaining unauthorized access to third-party systems" (a Chinese state-backed group used Claude Code against some 30 targets, including financial institutions), and that its safeguards against weapons misuse "have been and may in the future again be circumvented".

Most relevant for our readers is insurance: Anthropic warns that "with the scale of harm that could result from a catastrophic AI-related incident, no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity or expertise" to cover it.

Third, the opportunity. Why invest despite all the risks? Because the prize is large (knowledge work is worth more than $40tn a year, Anthropic says), customers already pay (every Fortune 10 company and the ten largest banks spend at least $100,000 a year, and existing customers spend roughly nine times more than a year ago), and the risks are part of the pitch: safety, Anthropic argues, has become one of the "decisive criteria" for enterprises choosing an AI partner. For our readers, the most interesting example is perhaps AIG: in its Lexington middle-market property business, underwriting with Claude cut time-to-quote by 55%, while quoted submissions rose 30% and bound submissions 40%.

So how did the market read all this? The same week, Hyperscalers released their Q3 results, and despite tremendous growth, investors are mostly focusing on the cash at hand:

Strong growth was not enough: Google Cloud doubled, yet Alphabet lost >$500bn in a single session as its cash flow turned negative. Investors are getting pickier in both public and private markets: OpenAI had to cut its latest funding target by $10bn, at a valuation already trimmed to $1.25tn. And insurers sit on both sides: US life insurers bought more than $740m of CoreWeave debt backed by Meta, just as underwriters are told AI's worst case may be beyond them.

Anthropic's filing does not disclose a share price or valuation, although rumours of a $2tn valuation persist.

Cyber, continued

Data centres on the ballot

Agents at the till

Insurance & finance

Work

Chips and China

As we head into AI lab IPO season, it might be a good moment to spend a minute trying to understand how, why and to what end those companies work. Karen Hao is perhaps one of the most interesting voices in the sphere, and I would strongly recommend her appearance on Bloomberg’s The Mishal Husain Show: Karen Hao on the Reckless Race for Global AI Power

Karen Hao, a former Silicon Valley insider turned journalist, deconstructs how tech giants have become modern “empires”, consolidating unprecedented power while often ignoring the human and environmental costs of their growth. By highlighting the exploitation of “data workers” and challenging the industry's “growth-at-all-costs” mentality, Hao invites us to consider a more sustainable, human-centred future, all while explaining why she personally chooses to live without these tools to preserve her own agency. If you want to understand the true price behind the technology defining our era, her view is an essential starting point.

She also wrote a book: Empire of AI — recommended reading.

Both of this week's stories point to the same gap: companies need cover for damage done to them by outside agents (Black October's hackers), but also for damage done by their own. Harbor, whose $10m seed round we mentioned briefly last edition, focuses on the second.

Out of stealth since mid-October, the San Francisco startup, founded by four engineers (ex-McKinsey, Citadel and Morgan Stanley; Stanford and MIT), helps companies understand, reduce and insure the risks of putting AI agents into production. Its first customers are banks and insurers in the US, Europe and Latin America.

The process has three steps. First, Harbor finds every agent a company runs. Second, it scores each one on operational, security and regulatory risk, by reading its activity logs, attacking it in a sandbox and, when needed, reviewing its code and prompts. Third, it adds guardrails that check every action before the agent executes it, translates what is left into an expected loss "in dollars, per year", and insures it: hallucinations, wrong actions, bias, broken contracts, data leaks and IP infringement.

So is Harbor an insurer? Not quite. Legally, it is a licensed insurance agent and surplus lines broker, and its policies are issued by surplus lines carriers. It aims to work as an MGA: pricing with its own risk models and distributing the product itself, while a carrier provides the "paper" (i.e. the balance sheet that pays the claims). In late October Harbor was in London lining up that capacity, with talks close to being finalised; the partners have not been named. Harbor's real asset is therefore the data on how agents fail and the ability to help businesses mitigate the risk.

Have a good week,